Topsy-Turvy: integrating a global view into sequence-based PPI prediction

Area of Science:

• Computational biology
• Bioinformatics
• Systems biology

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions.
• Existing computational methods for PPI prediction are either sequence-based ('bottom-up') or network-based ('top-down').
• Integrating global network insights into sequence-based prediction remains a challenge.

Purpose of the Study:

• To introduce Topsy-Turvy, a novel deep-learning method for PPI prediction.
• To develop a hybrid model (TT-Hybrid) for enhanced prediction accuracy in species with existing PPI data.
• To enable accurate, interpretable, and genome-scale PPI prediction, particularly for non-model organisms.

Main Methods:

• Topsy-Turvy utilizes a sequence-based, multi-scale deep-learning architecture.
• It employs a transfer-learning approach during training, incorporating global and molecular-level interaction patterns.
• TT-Hybrid integrates Topsy-Turvy with a network-based link prediction model.

Main Results:

• Topsy-Turvy achieves state-of-the-art performance in cross-species PPI prediction.
• TT-Hybrid outperforms constituent models and other methods for both well- and sparsely-characterized proteins.
• Both methods demonstrate genome-scale feasibility and scalability, outperforming methods like AlphaFold-Multimer.

Conclusions:

• Topsy-Turvy and TT-Hybrid offer accurate, generalizable, and scalable solutions for PPI prediction.
• These methods facilitate comprehensive mapping of protein interactions in both model and non-model organisms.
• The approach unlocks new possibilities for understanding protein organization and function at a genome-wide scale.